Architecture-of-Neural-Networks

Tasks for Architecture of Neural Networks Course at ITMO University

---

Index

1. Introduction to Pytorch and NN
2. CNN Basics
3. Image segmentation with U-Net
4. Deep Convolutional Generative Adversarial Network architecture

---

Introduction to Pytorch and NN

The lab file includes :

• PyTorch introductory assignments
• Feedforward Neural Network construction assignment and Training FNN

CNN Basics

This task is devoted to creating convolutional neural network from scratch, using only pure Python and NumPy package.

Image segmentation with U-Net

The task is devoted to implement Image segmentation with U-Net. The task uses the Oxford-IIIT Pet Dataset, created by Parkhi et al. The dataset consists of images of 37 pet breeds, with 200 images per breed (~100 each in the train and test split). Each pixel of an animal image corresponds to one of three classes: a pixel belonging to a pet; a pixel located at the border of a pet; all other pixels.

Deep Convolutional Generative Adversarial Network architecture

The task is devoted to implement the Deep Convolutional Generative Adversarial Network architecture (DCGAN). This task is based on the paper Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks. This paper by Alec Radford, Luke Metz, and Soumith Chintala was released in 2016 and has become the baseline for many Convolutional GAN architectures in deep learning. The goal is to learn about the DCGAN architecture from the paper and to implement the model using PyTorch deep learning framework. Through the task, we were asked to work with the CIFAR10 dataset and try to train models for generating similar images.

The generated images can be shown in the folder.